A Weighted Similarity Measure for k-Nearest Neighbors Algorithm

Abstract

One of the most important problems in machine learning, which has gained importance in recent years, is classification. The k-nearest neighbors (kNN) algorithm is widely used in classification problem because it is a simple and effective method. However, there are several factors affecting the performance of kNN algorithm. One of them is determining an appropriate proximity (distance or similarity) measure. Although the Euclidean distance is often used as a proximity measure in the application of the kNN, studies show that the use of different proximity measures can improve the performance of the kNN. In this study, we propose the Weighted Similarity k-Nearest Neighbors algorithm (WS-kNN) which use a weighted similarity as proximity measure in the kNN algorithm. Firstly, it calculates the weight of each attribute and similarity between the instances in the dataset. And then, it weights similarities by attribute weights and creates a weighted similarity matrix to use as proximity measure. The proposed algorithm is compared with the classical kNN method based on the Euclidean distance. To verify the performance of our algorithm, experiments are made on 10 different real-life datasets from the UCI (UC Irvine Machine Learning Repository) by classification accuracy. Experimental results show that the proposed WS-kNN algorithm can achieve comparative classification accuracy. For some datasets, this new algorithm gives highly good results. In addition, we demonstrated that the use of different proximity measures can affect the classification accuracy of kNN algorithm.